Valuation guides

Customer proxy instruments explained

What proxy models of liabilities and assets are, where they come from, worked examples and how a proxy is valued on today's market.

Definition

A proxy instrument is a position whose value the owner's own system has already reduced to a simple function of a few drivers, such as interest rates, an equity index, credit spreads or lapse rates. Insurers build them for liabilities (and sometimes assets) that take hours to revalue in their actuarial models, so that thousands of scenarios can be valued in seconds.

A proxy comes in one of three forms, or a mix of them. A delta-gamma proxy gives first and second derivatives in each driver, and cross terms between pairs. A polynomial proxy, such as a fitted quadratic or a higher-degree fit from least-squares Monte Carlo, gives coefficients on powers and products of the drivers. A grid proxy, such as a price ladder, gives the value at points along one or more drivers and is interpolated between them.

History

Delta-gamma approximations of portfolio value were used for value at risk from the mid-1990s, when full revaluation of large books was too slow. Insurers took up proxy modelling in earnest for Solvency II in the early 2010s, when the one-year value at risk of a liability needed many thousands of nested valuations. Curve fitting and least-squares Monte Carlo, a regression on simulated values in the spirit of Longstaff and Schwartz (2001), became the standard ways to build them.

Examples

A delta-gamma liability

An annuity book is worth −1,250 million pounds to the insurer. Its proxy has a delta of 950,000 pounds per basis point of the 10-year rate and a gamma of −1,800 pounds per basis point squared.

If the 10-year rate rises 50bp, the P&L is 950,000 × 50 + ½ × (−1,800) × 50² = 47.5 million − 2.25 million = 45.25 million pounds. The liability is then worth −1,204.75 million pounds.

A fitted cubic in the equity move

A least-squares Monte Carlo proxy for a guarantee is f = 2.0 z − 0.3 z² + 0.05 z³ million pounds, where z is the equity move in units of 15%, held within ±3.

If equities fall 30%, z = −2 and f = −4.0 − 1.2 − 0.4 = −5.6 million pounds.

A lapse ladder

A ladder gives the book's P&L at lapse moves of −50%, 0 and +50% (relative to the expected lapse rate) as −40 million, 0 and +25 million pounds.

At a +20% lapse move, linear interpolation gives 25 million × 20 / 50 = 10 million pounds.

Variations

Delta-gamma proxy
Deltas, gammas and cross gammas in any number of drivers.
Polynomial proxy
Any number of terms and any degree, on monomial, Legendre, Chebyshev or Hermite polynomials, with each driver centred, scaled and clipped.
Grid proxy
Ladders, surfaces and grids in more drivers, interpolated linearly, by monotone cubic (one axis) or cubic, often with a delta-gamma block for the drivers on no grid.
Mixed proxy
Any sum of delta-gamma, polynomial and grid blocks.

Valuation methods

Each driver is defined once: one market factor (a zero rate at a tenor, an FX rate, a volatility, an underlying price) or a weighted combination of them, such as a curve's level or slope, or a client factor such as lapse. Its move is the factor's level today against its level when the proxy was calibrated, in the driver's own convention (basis points, percent, log or absolute), times a scale.

The proxy is evaluated at those moves and its blocks are summed. A proxy built some months ago is therefore valued where today's market has moved it to, and a scenario's P&L is the proxy at the scenario's moves less the proxy at today's. Moves beyond the range the proxy was fitted on are held at the edge unless the driver says to extrapolate, because a polynomial is unbounded outside its fit.

The calculator cannot revalue the instrument itself. Its only check is against validation points from the client's own system, on which it reports the proxy's RMS and largest error.

Delta-gamma block

∑iδixi+12∑iγixi2+∑a<bcab xaxb\sum_i \delta_i x_i + \tfrac12 \sum_i \gamma_i x_i^2 + \sum_{a<b} c_{ab}\, x_a x_b

Polynomial block

c0+∑kck∏iBpki(zi),zi=clip⁡ ⁣(xi−misi)c_0 + \sum_k c_k \prod_i B_{p_{ki}}(z_i), \qquad z_i = \operatorname{clip}\!\left(\frac{x_i - m_i}{s_i}\right)

Driver move

x=s∑kwk m(Lk,Lk0)x = s \sum_k w_k\, m\big(L_k, L_k^{0}\big)

x is a driver's move, δ, γ and c the deltas, gammas and cross terms, B_p the basis polynomial of degree p, m and s a driver's centre and scale, w_k the weights on market factors with levels L_k today and L_k⁰ at calibration.

Models on RiskAtoms

InstrumentPricing models
Delta-gamma proxy
Customer proxy (delta-gamma, polynomial and grid blocks)
Polynomial proxy (fitted quadratic and higher)
Customer proxy (delta-gamma, polynomial and grid blocks)
Grid proxy (price ladders and surfaces)
Customer proxy (delta-gamma, polynomial and grid blocks)
Mixed proxy (any blocks)
Customer proxy (delta-gamma, polynomial and grid blocks)